* feat: add Caltech101 dataset

* feat: rename as Caltech101Silhouettes and split in train, val and test

When tensorpack was installed by wheel the platform was already evaluated
and thus python-prctl was installed, even though it is Linux-specific.
Installing from source package or sources was OK.

Instead use a platform specifier that is evaluated while installing the wheel.

still use `tf.get_variable` because `tf.Variable` sometimes takes name_scope rather than variable_scope